AgentSkillsCN

Issue

问题

SKILL.md

PM: Open Issue

When to Use

When you discover a problem that needs engineering attention.

PM Role

Open the issue. Don't solve it.

  1. Report symptoms — What's broken? What's the impact?
  2. Set urgency — P0/P1/P2 based on impact
  3. Define acceptance — What does "fixed" look like? (outcome-based)
  4. Hand off — Engineers investigate root cause and propose fix

Issue Template

markdown
---
to: {engineer}
type: issue
priority: P{0|1|2}
created: {timestamp}
---

# {Short problem description}

## Expected

{What should happen}

## Actual

{What's happening instead}

## Impact

{Who/what is affected, severity}

## Acceptance Criteria

- [ ] {Outcome 1}
- [ ] {Outcome 2}

Core Rule

PM raises problems. Eng solves them.

  • PM states: expected vs actual
  • PM defines: what "fixed" looks like
  • Eng investigates: why it broke
  • Eng proposes: how to fix it

After Handoff

  1. Stakeholder in RCA — Engineers lead, PM participates
  2. Verify understanding — Ask until you understand:
    • How exactly the problem happened
    • How exactly the proposed solution resolves it
  3. Don't accept hand-wavy fixes — "We'll refactor" is not a root cause

Anti-Patterns

WrongRight
PM writes root cause analysisEngineer writes, PM reviews
PM proposes implementationPM defines outcome, engineer proposes how
"Fix cn sync" (vague)"27 messages reach input.md" (outcome)
Accepting fix without understandingAsk questions until it clicks

MCA

The minimum issue contains:

  1. What's broken (symptoms)
  2. How bad (impact/priority)
  3. What success looks like (acceptance)

Everything else is engineering's job.


Asking Humans for Decisions

When escalating to or asking a human for a decision:

Know your MCA. Do it. Only ask if genuinely uncertain.

Wrong: "What should I do?" ❌ Wrong: "Should I do X or Y?" ❌ Wrong: "My MCA is X. Agree?" (still asking permission) ✅ Right: Just do it. ✅ Right (if uncertain): "My MCA is X because [reason], but I'm uncertain about [specific thing]."

Default behavior: Act on your MCA. Human will correct if needed.

Only ask when:

  • Stakes are high (irreversible, external-facing, security-sensitive)
  • You're genuinely uncertain and can articulate why
  • You lack context you can't get yourself

Example (certain):

"Sigma delivered P1 fix. Reviewed, looks sound. Merging now." (Then merge. Don't wait for approval.)

Example (uncertain):

"Option A is safer, Option B is faster. I'd lean A but this is your risk tolerance call." (Genuine uncertainty — asking is appropriate.)

If you don't know enough to have an MCA, say what's missing:

"Can't form an MCA — I need to know [X] first."